Presentation of the EURODELTA III intercomparison exercise ......B. Bessagnet et al.: Presentation of the EURODELTA III intercomparison exercise 12669 al., 2012), models clearly tend

Atmos. Chem. Phys., 16, 12667–12701, 2016www.atmos-chem-phys.net/16/12667/2016/doi:10.5194/acp-16-12667-2016© Author(s) 2016. CC Attribution 3.0 License.

Presentation of the EURODELTA III intercomparison exercise –evaluation of the chemistry transport models’ performance oncriteria pollutants and joint analysis with meteorologyBertrand Bessagnet1, Guido Pirovano2, Mihaela Mircea3, Cornelius Cuvelier4, Armin Aulinger5, Giuseppe Calori6,Giancarlo Ciarelli7, Astrid Manders8, Rainer Stern9, Svetlana Tsyro10, Marta García Vivanco11, Philippe Thunis12,Maria-Teresa Pay13, Augustin Colette1, Florian Couvidat1, Frédérik Meleux1, Laurence Rouïl1, Anthony Ung1,Sebnem Aksoyoglu7, José María Baldasano13, Johannes Bieser5, Gino Briganti3, Andrea Cappelletti3,Massimo D’Isidoro3, Sandro Finardi6, Richard Kranenburg8, Camillo Silibello6, Claudio Carnevale14, Wenche Aas15,Jean-Charles Dupont16, Hilde Fagerli10, Lucia Gonzalez17, Laurent Menut18, André S. H. Prévôt7, Pete Roberts17,and Les White191INERIS, National Institute for Industrial Environment and Risks, Parc Technologique ALATA,60550 Verneuil-en-Halatte, France2RSE S.p.A., via Rubattino 54, 20134 Milan, Italy3ENEA, Italian National Agency for New Technologies, Energy and Sustainable Economic Development (ENEA),Via Martiri di Monte Sole 4, 40129 Bologna, Italy4Ex European Commission, Joint Research Centre JRC Institute for Environment and Sustainability, 21020 Ispra (Va), Italy5HZG, Helmholtz-Zentrum Geesthacht, Institute for Coastal Research, Max-Planck-Straße 1, 21502 Geesthacht, Germany6ARIANET Srl, Via Gilino n.9 20128, Milan, Italy7PSI, LAC, Paul Scherrer Institute, 5232 Villigen PSI, Switzerland8TNO, Dept. Climate, Air and Sustainability, P.O. Box 80015, 3508 TA Utrecht, the Netherlands9Freie Universität Berlin, Institut für Meteorologie Troposphärische Umweltforschung Carl-Heinrich-Becker Weg 6–10,12165 Berlin, Germany10Climate Modelling and Air Pollution Division, Research and Development Department, Norwegian MeteorologicalInstitute (MET Norway) P.O. Box 43, Blindern, 0313 Oslo, Norway11CIEMAT, Atmospheric Pollution Unit, Avda. Complutense, 22, 28040 Madrid, Spain12European Commission, Joint Research Centre JRC Institute for Environment and Sustainability 21020 Ispra (Va), Italy13BSC, Barcelona Supercomputing Center, Centro Nacional de Supercomputación, Nexus II Building, Jordi Girona, 29,08034 Barcelona, Spain14Department of Electronics for the Automation, University of Brescia, via Branze 38, 25123 Brescia, Italy15Norwegian Institute for Air Research (NILU), Box 100, 2027 Kjeller, Norway16Institut Pierre-Simon Laplace, CNRS-Ecole Polytechnique, 91128 Palaiseau, Paris, France17CONCAWE, Boulevard du Souverain 165, 1160 Brussels, Belgium18Laboratoire de Météorologie Dynamique, École Polytechnique, ENS, UPMC, CNRS, Institut Pierre-Simon Laplace,91128 Palaiseau, France19AERIS EUROPE Ltd., Strouds Church Lane, West Sussex RH17 7AY, UK

Correspondence to: Bertrand Bessagnet ([email protected])

Received: 20 September 2015 – Published in Atmos. Chem. Phys. Discuss.: 25 January 2016Revised: 2 September 2016 – Accepted: 19 September 2016 – Published: 12 October 2016

Published by Copernicus Publications on behalf of the European Geosciences Union.

12668 B. Bessagnet et al.: Presentation of the EURODELTA III intercomparison exercise

Abstract. The EURODELTA III exercise has facilitated acomprehensive intercomparison and evaluation of chemistrytransport model performances. Participating models per-formed calculations for four 1-month periods in different sea-sons in the years 2006 to 2009, allowing the influence ofdifferent meteorological conditions on model performancesto be evaluated. The exercise was performed with strict re-quirements for the input data, with few exceptions. As a con-sequence, most of differences in the outputs will be attributedto the differences in model formulations of chemical andphysical processes. The models were evaluated mainly forbackground rural stations in Europe. The performance wasassessed in terms of bias, root mean square error and cor-relation with respect to the concentrations of air pollutants(NO2, O3, SO2, PM10 and PM2.5), as well as key meteoro-logical variables. Though most of meteorological parameterswere prescribed, some variables like the planetary bound-ary layer (PBL) height and the vertical diffusion coefficientwere derived in the model preprocessors and can partly ex-plain the spread in model results. In general, the daytime PBLheight is underestimated by all models. The largest variabil-ity of predicted PBL is observed over the ocean and seas.For ozone, this study shows the importance of proper bound-ary conditions for accurate model calculations and then onthe regime of the gas and particle chemistry. The modelsshow similar and quite good performance for nitrogen diox-ide, whereas they struggle to accurately reproduce measuredsulfur dioxide concentrations (for which the agreement withobservations is the poorest). In general, the models providea close-to-observations map of particulate matter (PM2.5 andPM10) concentrations over Europe rather with correlationsin the range 0.4–0.7 and a systematic underestimation reach-ing −10 µg m−3 for PM10. The highest concentrations aremuch more underestimated, particularly in wintertime. Fur-ther evaluation of the mean diurnal cycles of PM revealsa general model tendency to overestimate the effect of thePBL height rise on PM levels in the morning, while the in-tensity of afternoon chemistry leads formation of secondaryspecies to be underestimated. This results in larger modelledPM diurnal variations than the observations for all seasons.The models tend to be too sensitive to the daily variation ofthe PBL. All in all, in most cases model performances aremore influenced by the model setup than the season. Thegood representation of temporal evolution of wind speed isthe most responsible for models’ skillfulness in reproducingthe daily variability of pollutant concentrations (e.g. the de-velopment of peak episodes), while the reconstruction of thePBL diurnal cycle seems to play a larger role in driving thecorresponding pollutant diurnal cycle and hence determinesthe presence of systematic positive and negative biases de-tectable on daily basis.

1 Introduction

The ongoing project EURODELTA has very successfully ex-tended the European Air Quality Modelling capability byproviding a forum in which modelling teams could shareexperiences in simulating technically interesting and policy-relevant problems. The joint exercises contribute to furtherimproving modelling techniques as well as quantifying andunderstanding the sources of model uncertainties related tothe parameterization of processes and the quality of inputdata. EURODELTA is now an activity contributing to the sci-entific work of the UNECE (United Nations Economic Com-mission for Europe) Task Force on Measurement and Mod-elling (TFMM) under the Convention on Long-range Trans-boundary Air Pollution (CLRTAP). The TFMM was estab-lished in 2000 to provide a forum for the parties, the EMEP(European Monitoring and Evaluation Programme) centresand other international organizations for scientific discus-sions to evaluate measurements and modelling and to furtherdevelop working methods and tools. These are used for pol-icy studies in support of the Gothenburg Protocol signed in1999, which is a multi-pollutant protocol of the conventiondesigned to reduce acidification, eutrophication and ground-level ozone by setting emission ceilings for sulfur dioxide,nitrogen oxides, volatile organic compounds, fine particulatematter and ammonia.

In 2004, EURODELTA I (van Loon et al., 2007) examinedthe common performance of the chemistry transport mod-els (CTMs) in predicting recent (2000) and future (2020) airquality in Europe using the concept of a model ensemble tomeasure robustness of predictions. The spread of model pre-dictions about the ensemble mean gave a measure of uncer-tainty for each predicted value. In a 2020 world, the effectof making emission reductions for key pollutants in specificgeographic areas was investigated. The pollutants were NOx(nitrogen dioxide), SO2 (sulfur dioxide), VOC (volatile or-ganic compound), PM (particulate matter as PM10 and PM2.5for particle diameters below 10 and 2.5 µm, respectively) andNH3 (ammonia). The countries were France, Germany andItaly. The effect of reducing NOx and SOx in sea areas wasalso investigated. Source–receptor relationships used in in-tegrated assessment (IA) modelling were derived for all themodels and compared to assess how model choice might af-fect this key input. EURODELTA II (Thunis et al., 2008)built on this project by taking a closer look at how the dif-ferent models represent the effect on pollutant impacts on aEuropean scale by applying emission reductions to individ-ual emission sectors.

In the recent literature, several intercomparison and eval-uation exercises of regional-scale chemistry transport mod-els for PM have been reported: McKeen et al. (2007), vanLoon et al. (2007), Vautard et al. (2007, 2009), Hayami etal. (2008), Stern et al. (2008), Smyth et al. (2009), Solazzo etal. (2012), Pernigotti et al. (2013) and Prank et al. (2016).In one of the most recent exercises, AQMEII (Solazzo et

Atmos. Chem. Phys., 16, 12667–12701, 2016 www.atmos-chem-phys.net/16/12667/2016/

B. Bessagnet et al.: Presentation of the EURODELTA III intercomparison exercise 12669

al., 2012), models clearly tend to underestimate PM10 back-ground concentrations in US and EU regions. Model resultsfor PM2.5 concentrations showed better performances butlarge uncertainty remained certainly due to the simulation ofsecondary organic aerosols. Prank et al. (2016) stressed theproblems of emission underestimates to explain the modeldiscrepancies.

The new EURODELTA III (ED-III) exercise was designedto exploit and interpret intensive measurement campaignscarried out by EMEP (Aas et al., 2012). As far as possible,the models have been run in ED-III with the same input data(emissions, meteorology, boundary conditions) and over thesame domain (domain extension and resolution). This distin-guishes the study from other model intercomparisons. ED-IIIfocused on four EMEP intensive measurement periods:

– 1–30 June 2006

– 8 January–4 February 2007

– 17 September–15 October 2008

– 25 February–26 March 2009

The four different periods, within a rather limited numberof years, allowed the influence of different meteorologicalconditions on model performance to be evaluated. The listof modelling teams participating in the ED-III is reported inTable 1. FUB ran two of the four periods. The ED-III frame-work (emissions, model configurations) was also used to as-sess the impact of the horizontal resolution on the perfor-mance of air quality models (Schaap et al., 2015).

The ED-III exercise allowed a very comprehensive inter-comparison and evaluation of chemistry transport model per-formance with a joint analysis of some meteorological vari-ables to be made. A first evaluation of the 2009 campaignwith an interim version of models was published in Bessag-net et al. (2014). Moreover, the selected periods coincidewith EMEP intensive measurement periods so that an ex-tended set of observational data was available. Therefore, inaddition to EMEP operational monitoring data, size disag-gregated (in PM2.5 and PM10) aerosol data and hourly mea-surements for studying diurnal cycles have been used. Ad-ditional AirBase data (Mol and de Leeuw, 2005) were usedto evaluate the impact of meteorology on air pollutant con-centrations. Finally, the exercise was performed under strictrequirements (with some exceptions) concerning the inputdata. As a consequence, most of differences in the outputswill be attributed to the simulation of chemical and physicalprocesses. The objective of this paper is 2-fold: (i) to presentthe exercise, the input data and the participating models, and(ii) to analyse the behaviour of models in the four campaignsfocusing on the criteria pollutants PM10, PM2.5, O3, NO2 andSO2 as defined in the EU directive on air quality 2008/50/EC(EC, 2008) and relevant meteorological variables. Comple-mentary analyses of deposition fluxes and PM composition

data at high temporal resolution will be discussed in com-panion papers in order to better understand the behaviour ofmodels.

2 Description of models

2.1 Overall description of models

The models are synthetically described in Tables 2 and 3.All the models were run on the same domain at 0.25◦0.25◦

resolution in longitude and latitude except CMAQ. CMAQsimulations were performed on a Lambert-conformal conicprojection with the standard parallels at 30 and 60◦ and agrid of 112× 106 cells of size 24 km× 24 km. The results ofthe CMAQ simulations were interpolated to the prescribedEURODELTA grid.

Participants delivered both air concentrations and meteo-rological parameters. Most of variables were delivered on anhourly basis, while dry and wet deposition fluxes were pro-vided on a daily basis. The output species include, amongothers, O3, NO2 and SO2, total PM mass concentrations bothin 10 and 2.5 µm fractions (PM10 and PM2.5). Secondary in-organic aerosols such as ammonium (NH+4 ), sulfate (SO

2−4 )

and nitrate (NO−3 ) and other PM components relevant forthe analysis as well as wet deposition of sulfur and nitro-gen compounds were also collected and will be used in com-panion papers. The delivered air concentrations should ap-proximately correspond to the standard measurement height(typically 3 m) and were directly derived from the first modellayer, except for LOTOS-EUROS and EMEP that correctedthe concentrations from the first layer to be representative of3 m concentrations. The PM2.5 and PM10 concentrations arecalculated as follows in each model:

PMxx =PPMxx +∣∣∣SO2−4 ∣∣∣

xx+∣∣NO−3 ∣∣xx + ∣∣NH+4 ∣∣xx

+ |SOA|xx + |Dust|xx + |Sea Salts|xx ,

where xx = 2.5 or 10 µm; PPM stands for primary particu-late matter and includes elemental carbon, primary organicaerosol and primary non-carbonaceous aerosol; SOA repre-sents secondary organic aerosol; and sea salt and dust repre-sent the contribution of the corresponding natural processesmainly controlled by the wind speed.

The participating models differ in the availability of PMcomponents and formation routes. For instance, EMEP,LOTOS-EUROS and RCG contain coarse-mode nitrate for-mation (produced by reaction of nitric acid with sea salt anddust), whereas the others do not. In CMAQ, additional an-thropogenic dust is calculated as 90 % of unspecified PMcoarse emissions and attributed to fugitive dust (Binkowskyand Roselle, 2003). CAMx did not activate the parameteriza-tion of sea salt in this exercise.

Based on the setup of models and completeness ofdatasets, an ensemble called ENS has been built based on

www.atmos-chem-phys.net/16/12667/2016/ Atmos. Chem. Phys., 16, 12667–12701, 2016


Table 1. Models involved in the study.

Teams Models with Model acronym Simulatedreferences in this study periods

PSI/RSE CAMx (ENVIRON, 2011) CAMx 2006, 2007, 2008, 2009INERIS CHIMERE (Menut et al., 2013) CHIM 2006, 2007, 2008, 2009HZG CMAQ (Byun and Schere, 2006; Matthias et al., 2008) CMAQ 2006, 2007, 2008, 2009MSC-W – Met.NO EMEP (Simpson et al., 2012) EMEP 2006, 2007, 2008, 2009TNO LOTOS-EUROS (Sauter et al., 2014) LOTO 2006, 2007, 2008, 2009ENEA/ARIANET MINNI (ARIANET, 2004) MINNI 2006, 2007, 2008, 2009FUB RCG (Stern et al., 2006) RCG 2008, 2009

mean values of model outputs. To compare the behaviour ofmodels for all pollutants and campaigns, only CHIMERE,MINNI, LOTOS-EUROS and EMEP constitute the ensem-ble. CAMx, CMAQ and RCG were not included in the en-semble for three reasons: (i) CAMx did not account for seasalt leading to very different PM patterns over the oceans andseas, (ii) CMAQ used a different meteorology and (iii) RCGdid not cover the four campaigns.

2.2 PBL height and mixing in models

2.2.1 CAMx

In ED-III, the planetary boundary layer (PBL) was directlytaken from the ECMWF IFS data (Integrated Forecast Sys-tem of the European Centre for Medium-Range WeatherForecasts). The PBL height was then used by the CAMxpreprocessor to derive Kz profiles. For ED-III, the O’Brienscheme (1970) has been used to deriveKz profiles as Eq. (1):

Kz =KA+(z− zA)

2

(zA− zB)2{

KB −KA+ (z− zB)

(KB ′ + 2

KB −KA

zA− zB

)}, (1)

where Kz is a value of KA at the height of the atmosphericboundary layer zA, and KB at the height of the surface layerzB , the so-called constant-flux layer. Minimum Kz valueshave been set to 1 m2 s−1. Any values of Kz calculated be-low will be set to this value. By default, CAMx employs astandard K-theory approach for vertical diffusion to accountfor sub-grid-scale mixing layer to layer.

2.2.2 CHIMERE

In this study, the PBL is directly taken from the ECMWF IFSdata. Horizontal turbulent fluxes were not considered. Verti-cal turbulent mixing takes place only in the boundary layer.The formulation uses K diffusion following the parameteri-zation of Troen and Mahrt (1986), without a counter-gradientterm. In each model column, diffusivity Kz is calculated as

Eq. (2):

Kz = kwsz(

1−z

h

)1/3, (2)

where ws is a vertical velocity scale given by similarity for-mulae.

– In the stable case (surface sensible heat flux< 0): ws =u∗/ (

1+ 4.7z/L).

– In the unstable case: ws =(u3∗+ 2.8ew

3∗

)1/3,where e =max(0.1,z/h), L is the Monin–Obukhov length,w∗ is the convective velocity scale, u∗ the friction velocityand h the boundary layer height. The minimum value of Kzis assumed to be 0.01 m2 s−1.Kz and the wind speed were corrected in urban zones ac-

cording to Terrenoire et al. (2015) by applying a correctionfactor to limit the diffusion within the urban canopy, but thiscorrection has very little effect at this resolution.

2.2.3 CMAQ

The boundary layer height in COSMO is calculated with theturbulent kinetic energy (TKE) method (Doms et al., 2011).CMAQ directly used the PBL fields from COSMO.

In CMAQ, the vertical turbulent mixing is estimated us-ing the Asymmetric Convective Model scheme version 2(ACM2; Pleim, 2007a, b). The ACM2 replaces the simpleeddy viscosity (K-theory) scheme. The ACM2 scheme al-lows the non-local mixing, which means upward turbulentmixing from the surface across non-adjacent layers throughthe convective boundary layer. Pleim (2006) compared theeddy viscosity and the ACM2 schemes in CMAQ, findingthat the ACM2 scheme tends to predict larger concentrationsof secondary pollutants and smaller concentrations of pri-mary pollutants at the surface, and has a more well-mixedprofile in the PBL than the eddy viscosity scheme.

CMAQv5 also has an improved version of the minimumallowable vertical eddy diffusivity scheme. The new versioninterpolates between urban and non-urban land cover, allow-ing a larger minimum vertical diffusivity value for grid cells



Tabl

e2.

Synt

hetic

desc

ript

ion

ofm

odel

s(p

art1

).

EM

EP

CH

IME

RE

LO

TOS-

EU

RO

SR

CG

CM

AQ

MIN

NI

CA

Mx

Ver

sion

rv4.

1.3

Chi

mer

e201

3v1

.8v2

.1V

4.7.

1FA

RM

V3.

1.12

V5.

40

VE

RT

ICA

LM

OD

EL

STR

UC

TU

RE

Ver

tical

laye

rs20σ

9σ4

(3dy

nam

icla

yers

and

asu

rfac

ela

yer)

6fix

edte

rrai

n-fo

llow

ing

laye

rs30σ

16fix

edte

rrai

n-fo

llow

ing

laye

rs33σ

Ver

tical

exte

nt(h

Paor

m)

100

hPa

500

hPa

3500

m30

00m

100

hPa

1000

0m

8000

m

Firs

tlay

erde

pth

90m

20m

25m

25m

42m

40m

20m

Cor

rect

ion

offir

st-l

evel

conc

entr

atio

nY

esN

oY

esN

oN

oN

oN

o

NA

TU

RA

LE

MIS

SIO

NS

Bio

geni

cV

OC

Bas

edup

onm

aps

of11

5sp

ecie

sfr

omK

oe-

ble

and

Seuf

ert

(200

1)an

dho

urly

tem

pera

ture

and

light

;se

eSi

mps

onet

al.(

2012

)

ME

GA

Nm

odel

v2.0

4B

ased

upon

map

sof

115

spec

ies

from

Koe

ble

and

Seuf

ert

(200

1)an

dho

urly

tem

pera

ture

and

light

(Bel

t-m

anet

al.,

2013

)

Bas

edup

onm

aps

of11

5sp

ecie

sfr

omK

oebl

ean

dSe

ufer

t(2

001)

and

hour

lyte

mpe

ratu

rean

dlig

htus

ing

emis

sion

sfa

ctor

sof

Sim

p-so

net

al.(

1999

)

BE

IS3.

14em

issi

onin

-ve

ntor

y(V

ukov

ich

and

Pier

ce,2

002)

ME

GA

Nm

odel

v2.0

4M

EG

AN

mod

elv2

.1

Soil

NO

From

Sim

pson

etal

.(20

12)

ME

GA

Nm

odel

v2.0

4N

otus

edhe

reFr

omSi

mps

onet

al.(

1999

)B

EIS

3.1.

4M

EG

AN

v2.0

4M

EG

AN

mod

elv2

.1

Lig

htni

ngem

issi

ons

Clim

atol

ogic

alfie

lds,

Köh

lere

tal.

(199

5)N

oN

oN

oN

oN

oN

o

Sea

salt

Mon

ahan

(198

6),

Mar

tens

son

(200

3);s

eeT

syro

etal

.(20

11)

Mon

ahan

etal

.(19

86)

Mar

tens

son

etal

.(2

003)

,M

onah

anet

al.(

1986

)G

ong

etal

.(1

997)

,M

ona-

han

etal

.(19

86)

Zha

nget

al.

(200

5),

Cla

rke

etal

.(20

06)

Zha

nget

al.(

2005

)N

otus

ed

Win

dblo

wn

dust

Aft

erSi

mps

onet

al.(

2012

)N

oD

enie

rva

nde

rG

onet

al.(

2010

)L

oose

mor

ean

dH

unt

(200

0),

Cla

ibor

net

al.(

1998

)

No

Vau

tard

etal

.(20

05)

No

Roa

dtr

affic

susp

ensi

onD

enie

rva

nde

rG

onet

al.(

2010

)N

oN

oN

oN

oN

oN

o

LA

ND

USE

Lan

duse

data

base

CC

E/S

EI

for

Eur

ope,

else

whe

reG

LC

2000

GL

OB

CO

VE

R(2

4cl

asse

s)C

orin

eL

and

Cov

er20

00(1

3cl

asse

s)C

orin

eL

and

Cov

er20

00(1

3cl

asse

s)C

orin

eL

and

Cov

er20

06(4

4cl

asse

s)C

orin

eL

and

Cov

er20

06(2

2cl

asse

s)U

SGS

data

Res

olut

ion

Flex

ible

,C

CE

/SE

I∼

5km

Abo

ut30

0m

1/60×

1/60◦

1/60×

1/60◦

Abo

ut25

0m

Abo

ut25

0m

10m

in


12672 B. Bessagnet et al.: Presentation of the EURODELTA III intercomparison exerciseTable

3.Syntheticdescription

ofmodels

(part2).

EM

EP

CH

IME

RE

LO

TOS-E

UR

OS

RC

GC

MA

QM

INN

IC

AM

x

ME

TE

OR

OL

OG

Y

Driver

EC

MW

FIFS

EC

MW

FIFS

+urban

mixing

EC

MW

FIFS

EC

MW

FIFS+

observa-tions

CO

SMO

CL

ME

CM

WF

IFSE

CM

WF

IFS

Resolution

0.25◦×

0.25◦

0.25◦×

0.25◦

0.25◦×

0.25◦

0.25◦×

0.25◦

24km×

24km

(Lam

bertconform

alconic

projection)

0.25◦×

0.25◦

0.25◦×

0.25◦

PRO

CE

SSES

Advection

scheme

Bott(1989)

Van

Leer(1984)

Walcek

(2000)W

alcek(2000)

modified

byY

amartino

(2003)B

lackman

cubicpolynom

ials(Y

amartino,1993)

Blackm

ancubic

polynomials

(Ya-

martino,1993)

Bott(1989)

Verticaldiffusion

Kz

approachfollow

ingO

’Brien

(1970)and

onJeričevič

etal.

(2010)for

stableand

neutralconditions

Kz

approachfollow

ingTroen

andM

art(1986);IFS

PBL

Kz

approach;IFS

PBL

Kz

approach;IFSPB

LA

CM

2PB

Lschem

e(Pleim

,2007a)

Kz

following

Lange

(1989);PBL

fromM

aul(1980)version

ofC

ar-son

(1973)algorithmfordaytim

e

Kz

approachfollow

ingO

’Brien

(1970);IFS

PBL

Dry

depositionschem

eR

esistanceapproach

forgases,V

enkatramand

Pleim(1999)

foraerosols;

Simpson

etal.(2012)

Resistance

approach,Em

-berson

(2000a,b)R

esistanceapproach;

DE

-PA

C3.11

forgases;

Van

Zanten

etal.

(2010)and

Zhang

etal.

(2001)for

aerosols

Resistance

approach;D

E-

PAC

module

Resistance

approach,V

enka-tram

andPleim

(1999)R

esistancem

odel(W

alcekand

Taylor,1986;Wesely,1989)

Resistance

model

forgases

(Zhang

etal.,2003)

andaerosols

(Zhang

etal.,2001)

Com

pensationpoints

No,butzero

NH

3deposition

overgrowing

cropsN

oO

nlyfor

NH

3(for

stom-

atal,external

leafsurface

andsoil=

0)

No

No

No

No

Stomatalresistance

DO

3SE-E

ME

P:E

mberson

etal.

(2000a,b),

Tuovinenet

al.(2004),

Simpson

etal.(2012)

Em

berson(2000a,b)

Em

berson(2000a,b)

Wesely

(1989)W

esely(1989)

Wesely

(1989)W

esely(1989)

Wetdeposition

ofgasesIn-cloud

andsub-cloud

scavengingcoefficients

In-cloudand

sub-cloudscavenging

coefficientsSub-cloud

scavengingcoef-

ficientpH

-dependentscavenging

coefficientsIn-cloud

andsub-cloud

scav-enging

which

dependson

Henry’s

lawconstants,disso-

ciationconstants

andcloud

waterpH

,Chang

etal.(1987)

In-cloudand

sub-cloudscaveng-

ingcoefficients

(EM

EP,2003)

In-cloudand

sub-cloudscaveng-

ingm

odelfor

gasesand

aerosols(Seinfeld

andPandis,1998)

Wetdeposition

ofparticlesIn-cloud

andsub-cloud

scavengingIn-cloud

andsub-cloud

scavengingSub-cloud

scavengingcoef-

ficientSub-cloud

scavengingcoef-

ficientsIn-cloud

andsub-cloud

scav-enging

In-cloudand

sub-cloudscaveng-

ingcoefficients

In-cloudand

sub-cloudscaveng-

ingm

odelfor

gasesand

aerosols(Seinfeld

andPandis,1998)

Gas

phasechem

istryE

mC

hem09

(Simpson

etal.,2012)

ME

LC

HIO

RT

NO

CB

M-IV

CB

M-IV

CB

-05w

ithchlorine

chem-

istryextensions

(Yarw

oodet

al.,2005)

SAPR

C99

(Carter,2000a,b)

CB

-05(Y

arwood

etal.,2005)

Cloud

chemistry

Aqueous

SO2

chemistry

Aqueous

SO2

chemistry

andpH

computation

No

Simplified

aqueousSO

2chem

istryA

queousSO

2chem

istry(W

alcekand

Taylor,1986)A

queousSO

2chem

istry(Seinfeld

andPandis,1998)

Aqueous

SO2

chemistry

RA

DM

-A

Q(C

hangetal.,1987)

Coarse

nitrateY

esN

oreactions

with

Ca

orN

abut

coarsenessm

ightexist

with

transferfrom

smallerparticles

Yes

Yes

No

No

No

Secondaryinorganic

equilibriumM

AR

S(B

inkowski

andShankar,1995)

ISOR

RO

PIA(N

eneset

al.,1999)ISO

RR

OPIA

v2ISO

RR

OPIA

ISOR

RO

PIAv1.7

ISOR

RO

PIAv1.7

(Nenes

etal.,

1998)ISO

RR

OPIA

(Nenes

etal.,1998)

SOA

formation

VB

S-NPA

S,Sim

psonet

al.(2012)A

fterB

essagnetet

al.(2009)B

asedon

Bergström

etal.(2012)

SOR

GA

Mm

odule(Schell

etal.,2001)SO

RG

AM

module

(Schellet

al.,2001)SO

RG

AM

module

(Schellet

al.,2001)

CA

Mx-V

BS

(betaversion)

(Koo

etal.,2014)

VB

SY

es,Bergström

etal.(2012),Sim

psonetal.(2012)

No

Yes,based

onB

ergströmet

al.(2012)N

oN

oN

oY

es,basedon

Koo

etal.(2014)

Aerosolm

odelB

ulkapproach

(2m

odes)8

bins(40

nmto

10µm

)B

ulkapproach

(2m

odes)B

ulkapproach

(2m

odes)A

ER

O5

(Carlton

etal.,2010),L

og-normal

approach(3

modes)

AE

RO

3(B

inkowski,

1999);3

modes:

Aitken,

accumulation,

coarse

Bulk

approach(2

modes)

Aerosolphysics

No

dynamics

Coagulation/

condensa-tion/nucleation

No

dynamics

No

dynamics

Coagulation/

condensation/nucleation

Coagulation/

condensation/nu-

cleationN

odynam

ics



that are primarily urban. Moreover, the minimum eddy diffu-sivity values were reduced from 0.5 to 0.01 m2 s−1, and from2.0 to 1.0 m2 s−1 for urban areas.

2.2.4 EMEP

The mixing height is calculated using a slightly modifiedRichardson number (RiB) following Jeričevič et al. (2010)and defined as the lowest height at which the RiB > 0.25. Fi-nally, the PBL is smoothed with a second-order Shapiro filterin space. The PBL height is not allowed to be less than 100 mor exceed 3000 m.

The initial calculation of the vertical exchange coefficientsis done using the Ri number and wind speed vertical gradi-ent for the whole domain. Then, Kz values within the PBLare recalculated based on Jeričevič et al. (2010) for stableand neutral conditions. For unstable situations, Kz is calcu-lated based on the similarity theory of Monin–Obukhov forthe surface layer, whereas Kz profiles from O’Brien (1970)are used for the PBL above the surface layer. For more de-tails, see Simpson et al. (2012).

2.2.5 LOTOS-EUROS

The first model layer is, by definition, the mixing layer,with height equal to the boundary layer height as given byECMWF. Horizontal diffusion is not used, but for verticalmixing the vertical diffusion coefficient is calculated accord-ing to Eq. (3):

Kz =κu∗

8(z/L), (3)

where κ is the von Karman constant; u∗ the friction velocity;8 the functions proposed by Businger (1971) for stable, neu-tral or unstable atmosphere; z the height and L the Monin–Obukhov length. The friction velocity is calculated depend-ing on the wind at reference height (10 m), the Businger func-tions and the roughness length per land use class. The verti-cal structure of LOTOS-EUROS is determined by the mix-ing layer height, with a shallow surface layer (25 m) to avoidmixing of near-surface emissions that is too fast and a secondlayer equal to the mixing layer as given by ECMWF.

2.2.6 MINNI

In MINNI, the friction velocity and Monin–Obukhov lengthare determined by using the Holtslag and van Ulden (1983)iterative scheme for unstable conditions and the Venka-tram (1980) iterative method for stable conditions. Micro-meteorological parameters over water are derived with theprofile method, using air–sea temperature difference (Hannaet al., 1985), with the needed roughness length, depending onwind speed, supplied by the Hosker (1974) parameterization.

During daytime, both convective and mechanical heightsare determined, keeping then the maximum value between

the two parameters. The convective height is calculated fol-lowing the Maul (1980) version of the Carson (1973) al-gorithm, essentially based on the heat conservation equa-tion. The mechanical mixing height is estimated by usingthe Venkatram (1980) algorithm. During nighttime, the bulkRichardson number method is applied (Sorensen, 1998), inwhich the height of the boundary layer is given by the small-est height at which the bulk Richardson number reaches thecritical value fixed to 0.25.

2.2.7 RCG

The mixing layer depth in the model is the height of thelayer closest to the input boundary layer height taken fromthe ECMWF IFS data. Vertical diffusion parameters for sta-ble and unstable conditions are derived using the Monin–Obukhov similarity theory for the diabatic surface layer. Thefriction velocity and Monin–Obukhov length are calculatediteratively depending on the 10 m wind, the stability correc-tion factors and the roughness length determined from landuse.

3 Input data

3.1 Anthropogenic emissions

The first step in the emission preparation was to calculatethe spatial pattern of emissions for the reference year 2007,which was selected because it was a key year for the TNO-MACC inventory (Kuenen et al., 2011). The anthropogenicemission input was harmonized following the methodologydescribed in Terrenoire et al. (2015). The total emissionsper sector and country were then scaled to the correspond-ing year of the campaigns: 2006, 2007, 2008 and 2009.Emission categories are broken down into 11 classes calledSNAP (selected nomenclature for air pollutants): (1) pub-lic power stations, (2) residential and comm./inst. combus-tion, (3) industrial combustion, (4) production processes,(5) extraction and distribution fossil fuel, (6) solvents use,(7) road traffic, (8) other mobile sources (trains, shipping,aircraft, etc.), (9) waste treatment and (10) agriculture. Nat-ural emissions (11) were calculated by the models as set outin Sect. 3.2.

The gridded distribution of anthropogenic emissions wasprovided by INERIS and it was based on a merging of differ-ent databases from

– TNO 0.125◦× 0.0625◦ emissions for 2007 from MACC(Kuenen et al., 2011);

– EMEP 0.5◦× 0.5◦ emission inventory for 2009(Vestreng et al., 2007);

– emission data from the GAINS database (http://gains.iiasa.ac.at/gains).


http://gains.iiasa.ac.at/gainshttp://gains.iiasa.ac.at/gains


Emission regridding was based on INERIS expertise and per-formed by means of various proxies:

– population data coming from the EEA database mergedwith global data (from the Socioeconomic Data and Ap-plications Center http://sedac.ciesin.columbia.edu) tofill gaps in Europe;

– the US Geophysical Survey land use at 1 km resolution(http://www.usgs.gov/);

– French bottom-up emission data for wood combustionto derive a proxy based on population density;

– EPER data for industries. The EPER decision is basedon Article 15(3) of Council Directive 96/61/EC (EC,1996) concerning integrated pollution prevention andcontrol. EPER is a web-based register which enables thepublic to view data on emissions of 50 key pollutants towater and air from large and medium-sized industrialpoint sources in the European Union.

The TNO-MACC dataset provides two distinct datasets:(i) large point sources (LPSs) with the coordinates of stacksand (ii) surface emissions on a fine grid (0.125◦× 0.0625◦).In the gridding process, the first step consisted in summingup LPS emissions from the TNO-MACC emissions inven-tory for 2007 with surface emissions to obtain total emis-sions as in the EMEP inventory. LPSs were aggregated withsurface emissions because no data were available to calculateplume rise heights for point sources emissions. For the vari-ous SNAP sectors, the processing steps were the following:

– SNAP 2: The country emissions were regridded withcoefficients based on population density and Frenchbottom-up data; the methodology (Terrenoire et al.,2015) was extrapolated to all of Europe. For PM2.5emissions, the annual EMEP national totals werekept except in the following countries: Czech Re-public, Bosnia and Herzegovina, Belgium, Belarus,Spain, France, Croatia, Ireland, Lithuania, Luxem-bourg, Moldova, Republic of Macedonia, the Nether-lands, Turkey. For these countries, PM2.5 emissionsfrom GAINS were used as this database provideshigher numbers and certainly more realistic since woodburning is known to be underestimated in the EMEPdatabase (Denier van der Gon et al., 2015). Additionalfactors were applied on two Polish regions for bothPM2.5 and PM10 emissions. As a preliminary solution,domestic combustion emissions from provinces with ac-tive coal mines were multiplied by a factor of 8, whilethose in neighbouring provinces were adjusted by a fac-tor of 4 (Kiesewetter et al., 2015). The former activityin coal mine regions still leads to high emissions of PMdue to domestic uses of coal.

– SNAP 3,7,8,9,10: TNO-MACC emission spatial dis-tribution was used as a proxy to regrid EMEP0.5◦× 0.5◦ annual totals into the finer modelling grid.

– SNAP 1,4,5,6: EMEP 0.5◦× 0.5◦ emissions were re-gridded by using artificial area (or built-up area), exceptfor industries where EPER data were used.

For countries where TNO-MACC emissions were notavailable, the EMEP 0.5◦× 0.5◦ emissions were used (Ice-land, Liechtenstein, Malta and Asian countries) and regrid-ded with adequate proxies (artificial land use, EPER data forindustries).

The following emitted species were used in the mod-els: methane (this species comes from the TNO-MACC in-ventory), carbon monoxide, ammonia, sulfur oxides, non-methane volatile organic compounds (NMVOCs), nitrogenoxides, primary particulate matter.

Residential emissions of particulate matter are dominant inwintertime. In most countries, they come from wood-burningor coal uses. Germany, Sweden, Spain clearly have the low-est levels of PM2.5 emissions for this activity sector. Roma-nia, Poland and France have the highest levels of annual totalemissions per country (Terrenoire et al., 2015). For this activ-ity sector, the PM2.5 emissions by components are providedin the Supplement S8.

The time profiles are those used in Thunis et al. (2008).Three types of profiles were provided:

– Seasonal factors with one value per species, month, ac-tivity sector and country.

– Weekly factors with one value per species, day of theweek (Monday–Sunday), activity sector and country.

– Hourly factors with one value per hour (local time),species and activity sector.

The vertical injection profile in CTMs was prescribed ac-cording to Bieser et al. (2011) where industrial sectors andresidential heating were assigned in lower levels comparedto the lower vertical levels than other literature default pro-files (Mailler et al., 2013).

Since only PM2.5 and coarse PM emissions were pro-vided by EMEP, a PM speciation profile provided by IIASA(based on Klimont et al., 2013) was used to estimate the frac-tion of non-carbonaceous species, elemental carbon and or-ganic matter per activity sectors and country. Models usedtheir own split for NOx , SOx and NMVOC emissions. Theseemission inventories did not account for recent changes inthe way to measure semi-volatile organic compounds fromwood-burning emissions as discussed in Denier van der Gonet al. (2015).

The full emission dataset is available upon request toINERIS.

3.2 Natural emissions

3.2.1 Biogenic VOC emissions from vegetation

CHIMERE and MINNI used version 2.04 of the MEGANmodel, while CAMx used version 2.1 (Guenther et al., 2006,


http://sedac.ciesin.columbia.eduhttp://www.usgs.gov/


2012). The model of emissions of gases and aerosols fromnature (MEGAN) is a modelling framework for estimatingfluxes of biogenic compounds between terrestrial ecosystemsand the atmosphere using simple mechanistic algorithms toaccount for the major known processes controlling biogenicemissions. It is available as an offline code and has also beencoupled into land surface and atmospheric chemistry models.

EMEP, LOTOS-EUROS and RCG used parameterizationsderived from Simpson et al. (1999) for the temporal varia-tions according to temperature and light, with maps of treespecies from Koeble and Seufert (2001).

CMAQ used the BEIS (Biogenic Emission Inventory Sys-tem; Vukovich and Pierce, 2002) module developed by theUS EPA. BEIS estimates volatile organic compound (VOC)emissions from vegetation and nitric oxide (NO) and car-bon monoxide (CO) emissions from soils. Because of re-source limitations, recent BEIS development has been incor-porated into the Sparse Matrix Operational Kernel Emissions(SMOKE) system (available at https://www.cmascenter.org/smoke), so that the native version of BEIS is built within theSMOKE architecture.

3.2.2 Soil nitrogen monoxide (NO) emissions

CHIMERE and MINNI used version 2.04 and CAMx usedversion 2.1 of the MEGAN model to calculate the NO emis-sions. RCG used a parameterization of NO emissions de-scribed in Simpson et al. (1999). LOTOS-EUROS did not in-clude NO emissions in this simulation. CMAQ used the BEISmodule. The soil NO emission parameterization for EMEP isdescribed in Simpson et al. (2012).

3.2.3 Sea salt emissions

All models host very different schemes based on Monahan(1986) with some updates from Martensson et al. (2003) forLOTOS-EUROS, and Gong et al. (1997) for RCG. CMAQand MINNI used the Zhang et al. (2005) parameterizationand CAMx had no sea salt for this exercise due to uncertaintythat is too high in sea salt parameterization. EMEP used pa-rameterization from Monahan (1986) for larger sizes of seaspray and Martensson et al. (2003) for smaller sizes.

CMAQ also emits sea salt sulfate using a fraction of7.76 % of emitted sea salt split into the accumulation andcoarse modes.

3.2.4 NO emissions from lightning

The only model to describe NO emissions from lightning isthe EMEP model, following Köhler et al. (1995).

3.2.5 Wildfire emissions

Fire emissions were provided by the GFASv1.0 database(Kaiser et al., 2012) only for the 2006 campaign. The GlobalFire Assimilation System (GFASv1.0) calculates biomass

burning emissions by assimilating fire radiative power (FRP)observations from the MODIS instruments onboard the Terraand Aqua satellites. It corrects for gaps in the observations,which are mostly due to cloud cover, and filters spurious FRPobservations of volcanoes, gas flares and other industrial ac-tivities. For all models, the wildfire emissions were assignedin the whole PBL layer. Only the following species weretaken into account: CO, CH4, NOx , SO2, PM2.5, TPM (to-tal primary matter), OC (organic carbon) and EC (elementalcarbon).

3.2.6 Dust emissions

For CAMx, CHIMERE and CMAQ, no natural dust mod-ule is activated for this exercise. For these three models,natural dust only comes from the boundary conditions. ForEMEP, windblown dust parameterization is documented inSimpson et al. (2012), road dust calculations are includedin the calculations from Denier van der Gon et al. (2010).LOTOS-EUROS contains emission parameterizations forseveral sources of mineral dust (Schaap et al., 2009). Onlywind-blown dust, resulting from wind erosion of bare soils,was taken into account here, together with dust from bound-ary conditions. Other sources (agricultural activities, roaddust resuspension) were not activated in ED-III. In MINNI,dust emissions from local erosion and particle resuspension(Vautard et al., 2005) with attenuation in the presence of veg-etation from Zender et al. (2003) is activated in this exercise.RCG considers resuspension of mineral aerosol as a functionof friction velocity and the nature of soils. Two mechanismsare treated: direct release of small dust particles by the wind(Loosmore and Hunt, 2000), and indirect release by collisionwith bigger soil grains that are lifted by the wind but returnto the surface by sedimentation (saltation process from Clai-born et al., 1998).

3.3 Meteorology

All models except CMAQ and RCG share the same mete-orological dataset at 0.2◦ resolution based on ECMWF IFS(Integrated Forecast System) calculations.

Because of its importance for applications (e.g. in air pol-lution modelling), the boundary layer height, as diagnosed inthe ECMWF IFS model, was made available. The parameter-ization of the mixed layer (and entrainment) uses a boundarylayer height from an entraining parcel model. But, in order toget a continuous field, in neutral and stable situations the bulkRichardson method proposed by Troen and Mahrt (1986)is used as a diagnostic, independently of the turbulence pa-rameterization. Boundary layer height is defined as the levelwhere the bulk Richardson number, based on the differencebetween quantities of energy at that level and the lowestmodel level, reaches the critical value Ricr = 0.25.

For RCG, a different meteorological dataset was used.The 3-D data for wind, temperature, humidity and density


https://www.cmascenter.org/smokehttps://www.cmascenter.org/smoke


were produced employing a diagnostic meteorological anal-ysis system developed at Freie Universität (Berlin, Germany)and based on an optimum interpolation procedure on isen-tropic surfaces. The system takes into account all availableobserved synoptic surface and upper air data as well as to-pographical and land use information (Reimer and Scherer,1992). Rain data, cloud data and boundary layer heights wereretrieved from the IFS dataset. Boundary layer parameters asfriction velocity and Monin–Obukhov length were calculatedon-the-fly by applying standard boundary layer theory.

The CMAQ model used meteorological variables calcu-lated with the COSMO model in CLimate Mode (COSMO-CLM) version 4.8 CLM 11. The COSMO model is the non-hydrostatic operational weather prediction model appliedand further developed by the national weather services joinedin the COnsortium for SMall scale MOdeling (COSMO) de-scribed in Bettems (2015).

3.4 Boundary conditions

In this study, the MACC reanalysis was used as input datafor the boundary conditions (Inness et al., 2013; Benedettiet al., 2009). The MACC II project (Modelling AtmosphericComposition and Climate) established the core global andregional atmospheric environmental service delivered as acomponent of the Copernicus initiative (http://copernicus.eu/). The reanalysis production stream provides analysesand 1-day forecasts of global fields of O3, CO, NO2, SO2,HCHO, CO2, CH4 and aerosols. Other reactive gases areavailable from the coupled chemistry transport model. Thereanalysis covers the period 2003–2011 with a 1-monthspinup. It runs at approximately 78 km× 78 km horizontalresolution over 60 levels. The coupled chemistry transportmodel has the same 60 vertical levels and a horizontal resolu-tion of 1.125◦× 1.125◦. For aerosols only elemental carbon,organic carbon, dust and sulfate were used.

Stratospheric ozone fields from the MACC reanalysisagree with ozone sondes and ACE-FTS (atmospheric chem-istry experiment Fourier transform spectrometer) data within±10 % in most seasons and regions. In the troposphere,the reanalysis shows biases of −5 to +10 % with re-spect to ozone sondes and aircraft data in the extratropics,while larger negative biases are shown in the tropics. Area-averaged total column ozone agrees with ozone fields froma multi-sensor reanalysis dataset within a few percent. Foraerosols, the observed Aerosol Optical Depth (AOD) is as-similated in the model with a feedback on individual PMspecies (sea salt, dust, elemental carbon, organic carbon andsulfate). When available, the MACC reanalysis is comparedwith observations, the model acronym in the supporting ma-terial is MACCA.

4 Observation dataset and statistics

4.1 Air pollutant concentrations

The evaluation was carried out with the available EMEPstandard monitoring (Tørseth et al., 2012) and intensive pe-riod observations for 2006, 2007, 2008 and 2009 (Aas etal., 2012) on hourly and daily bases (see Supplement S8for the description of background sites). Elevated sites above1500 m in altitude have been excluded from the analysis. Themeasurements were downloaded from the EBAS database(http://ebas.nilu.no/). Additional AirBase data (Mol and deLeeuw, 2005) were used to evaluate the impact of meteorol-ogy on air pollutant concentrations in Sect. 7.2.

It is important to note that daily measurement for a day,N , is the averaged value between day N HH:00 and dayN+1 HH:00, with HH usually varying in the range [00, 09]in GMT. For most of the species, measurements on a dailyand hourly basis are not necessarily performed for the sameset of stations. Deposition and the PM composition are alsoavailable; the dataset will be detailed in the companion pa-pers.

4.2 Meteorology

4.2.1 Temperature and wind speed

The temperature, wind speed and precipitation measure-ments come from 2016 synoptic stations in Europe reportedby the European meteorological centres. The data are pro-vided on an hourly basis. The temperature is measured at2 m and the wind speed at 10 m. Some meteorological dataare also reported at some EMEP sites. At EMEP sites, dailyaccumulated measurements (e.g. precipitation) for a day Nrepresent the integral between dayN at HH:00 and dayN+1at HH:00, with HH usually varying in the range [00, 09] inGMT.

4.2.2 PBL height

The sounding data were extracted from the University ofWyoming database (http://weather.uwyo.edu/). For each siteand each day, two soundings are available at 00:00 and12:00 GMT. The provided meteorological parameters arepressure (hPa), the corresponding height above ground level(m), dew point temperature (◦C), relative humidity (%), mix-ing ratio (g kg−1), wind direction (degrees) and wind speed(expressed in knots and converted to m s−1 by applying theconversion factor 0.514), potential and virtual potential tem-perature (K). For the present study, data were extracted over77 stations in Europe. The boundary layer height is esti-mated using the calculation of the bulk Richardson num-ber profile and searching for the altitude where the criticalvalue of Ricr = 0.25 is reached. The analysis was limited tothe first 25 vertical points, roughly corresponding to an alti-tude of 5000 m above ground level. Since the boundary layer


http://copernicus.eu/http://copernicus.eu/http://ebas.nilu.no/http://weather.uwyo.edu/


Table 4. Error statistics used to evaluate model performance (M and O refer, respectively, to model and observation data, and N is thenumber of observations).

Mean bias(M̄ − Ō

)with M̄ = 1

N

N∑i=1

Mi and Ō = 1NN∑i=1

Oi

Normalized mean bias NMB=(M̄ − Ō

)/Ō

Mean bias MB=(M̄ − Ō

)Mean gross error MGE= 1

N

N∑i=1|Mi −Oi |

Standard deviation SDX =

√1N

N∑i=1

(Xi − X̄

)2 with X=O or MRoot mean square error RMSE=

√1N

N∑i=1

(Mi −Oi)2

Normalized root mean square error NMSE= 1M̄

√1N

N∑i=1

(Mi −Oi)2

Correlation coefficient R =

(N∑i=1

(Mi − M̄

)(Oi − Ō

))/(√ N∑i=1

(Mi − M̄

)2×

N∑i=1

(Oi − Ō

)2)

height is a concept valid only for convective periods, onlythe soundings of 12:00 GMT were analysed and used for themodel evaluation.

In addition to the previous PBL data, hourly heights ofthe atmospheric boundary layer were calculated from lidarmeasurements in a background site near Paris (SIRTA inPalaiseau, France). A new objective method for the determi-nation of the atmospheric boundary layer depths using rou-tine lidar measurements has been used (Pal et al., 2013).

4.3 Error statistics for the evaluation of modelperformances

The errors statistics considered in this report are presentedin Table 4. In the Supplement S0–S1 the performances of allmodels for the four campaigns are reported. For a given pol-lutant or meteorological variable, model performance is com-puted for a common set of stations (over the same commongeographic area). All maps of pollutant concentrations andmeteorological variables concerning individual models andensemble are provided in the Supplement (Sects. S2–S6).

For the analysis of the ensemble, a coefficient of variation(VAR) is defined as follows in Eq. (4):

VAR=1

CENS

√1M

∑m

(Cm−CENS)2, (4)

where Cm is the concentration of individual model mincluded in the ensemble (CHIMERE, LOTOS-EUROS,MINNI and EMEP),M is the number of models and CENS isthe ensemble mean concentration.

5 Evaluation of the meteorology

Some general features for each campaign can be provided;they are taken from the NOAA (National Oceanic and Atmo-spheric Administration) global analysis (https://www.ncdc.noaa.gov/sotc/global/).

June 2006 temperatures were above average everywherein Europe with low precipitation except in Balkan countriesand Spain compared to the 1961–1990 base period.

January 2007 was characterized by windy conditions inEurope with temperatures above the average everywhere ex-cept in Spain, where temperatures were close to the aver-age values. In the beginning of February, temperatures wereparticularly low in Scandinavia. Precipitation was low overthe Mediterranean basin but above the climate average, com-pared to the 1961–1990 period in the rest of Europe.

In September–October 2008, no clear general character-istics were recorded; this transition period was character-ized by slight negative temperature anomalies over the west-ern part of Europe, mainly France, the United Kingdom andnorth of Spain.

After some cold spells in the end of February, March 2009turned milder with average warmer temperatures comparedto the 1961–1990 base period. Precipitation was below aver-age in the western part of Europe and above average in thecentral and eastern parts of Europe.

5.1 The 2 m temperature

As summarized in the Supplement S0, the models usingECMWF data show comparable high temporal correlation


https://www.ncdc.noaa.gov/sotc/global/https://www.ncdc.noaa.gov/sotc/global/


Figure 1. Comparisons of observed vs. predicted meteorological variables (U10, T2M) for the 2009 campaign. Top left panel: mean diurnalcycle of the 10 m wind speed; top right panel: mean diurnal cycle of the 2 m temperature; bottom left panel: mean 10 m wind speed forCHIMERE; bottom right panel: mean 2 m temperature for CHIMERE. Some observations at EMEP stations are provided with colouredcircles over the maps. Red colour is assigned for values exceeding the colour scale.

coefficients based on hourly values over the whole domain(0.88


tiveness for inland sites. Moreover, errors in wind speed mea-surements are higher for low winds. For the lowest winds, thecomparison of the predicted diurnal cycle with observationsshows a larger positive bias at night than during the afternoon(Fig. 1), this behaviour could lead to an overestimation of theadvection process in the chemistry transport models.

Time correlations are better for models using ECMWFdata but all models exhibit low correlations over the Alpineregion (north of Italy, southeast of France, Switzerland andAustria). The RCG model shows higher correlation coeffi-cients over northern Europe (Finland and Sweden) for the2009 campaign.

5.3 PBL and mixing

As explained in Sect. 4.2, the observed PBL height was cal-culated at 12:00 GMT because of methodology hypotheses,except at the SIRTA site where hourly measurements areavailable for 2008 and 2009. All models have a negative bias.The lowest RMSEs are shown for CAMx and CHIMEREwhich use the ECMWF PBL; the biases are in the range−237 and −100 m for these two models. It is worth notingthat CAMx and CHIMERE exhibit exactly the same perfor-mance, while LOTOS-EUROS and EMEP, which adoptedIFS PBL too, show partially different performances. Somedifferences are attributed to different interpolation schemesand the use of minimum PBL values during nighttime asfor EMEP. The largest underestimation of the PBL heightis usually found for MINNI particularly for the 2006 cam-paign (up to −616 m) and EMEP (up to −451 m) and thecorrelation coefficients for these models are lower comparedto the others. CMAQ has the lowest bias for most of cam-paigns. Models using IFS PBL data showed the best per-formance for temporal correlation (see Supplement S0), themain discrepancies are observed for the 2006 campaign withseveral sites in Europe with negative correlations. The largestnegative biases are observed in the south of the domain; inthese regions CMAQ performs better. In some regions overthe Mediterranean basin, particularly in coastal areas, theMINNI’s PBL is sometimes strongly biased up to −1000 m.The obtained results suggest that either the Carlson algorithmor the micro-meteorological parameterization implementedby MINNI tends to underestimate the intensity of convection.

The spatial representation of the PBL for the 2009 cam-paign shows higher differences between the models mainlyover the ocean and seas where the coefficient of varia-tion reaches 40 % in some areas (Fig. 2). While LOTOS-EUROS, CHIMERE, RCG and CAMx use the PBL from IFSwith some differences in spatial and time interpolations, theother models use their own parameterizations discussed inSect. 2.2. The diurnal cycles displayed in Fig. 2 show thatMINNI simulates a higher PBL at night and a lower PBLduring daytime compared to ECMWF. The difference in theafternoon PBL is quite important over countries influencedby the ocean like Great Britain. CMAQ and EMEP simulate

the highest PBL at night over France and Great Britain. Thehourly times series at the SIRTA site confirm the underesti-mation of the ECMWF PBL but at this station, the negativebias of MINNI is of the same order of magnitude as thoseof the other models. The correlations based on hourly valuesare somewhat lower for CMAQ, EMEP, MINNI (below 0.50)compared to the models using ECMWF data.

The differences in treatment of advection and mixing asreported in Sect. 2.2 can lead to differences in the recon-struction of pollutant dispersion. Figure 3 shows the meancoefficient of variation of CO concentrations predicted by themodels sharing the same raw meteorology (IFS) for the 2006campaign. This pollutant can be considered a tracer with lowinfluences of deposition and chemistry processes, most of thedifferences in concentrations are related to transport and mix-ing. The figure clearly shows that mixing in emission areas,such as big cities, produces the highest differences exceed-ing 20 % of variations. The next highest coefficients of vari-ation are observed over the seas and ocean, which are relatedto the differences of PBL predicted by the models (Fig. 2);elsewhere, this coefficient remains below 10 %.

6 Overall model performance evaluation of criteriapollutants

6.1 Ozone

The model performances (Supplement S1) are very differentfrom campaign to campaign. Most of the models overesti-mate ozone concentrations in 2006, 2007 and 2008 (Fig. 4).Only the 2009 campaign shows a systematic underestimationof observed ozone concentrations from −5 to −16 µg m−3.The large positive bias in 2007 and negative in 2009 arelargely explained by the boundary conditions that are biased,respectively, by +8 and −20 µg m−3 (Supplement S1). Forthe positive bias in 2007, the boundary conditions cannot bethe sole reason, chemical processes play an important role.Correlations are similar for all models in the range 0.5–0.6,only CMAQ has lower correlations, on average. For the sum-mertime campaign in 2006 CHIMERE and CMAQ displaythe lowest correlation for daily averaged concentrations butCHIMERE has the lowest bias with EMEP. The low corre-lation for CMAQ and CHIMERE is due to the difficultiesin reproducing both spatial patterns and day-to-day varia-tions. For this campaign, most models underestimate con-centrations in the mountainous regions in Spain and over theAlps (Fig. 5). The models tend to overpredict ozone con-centrations in background stations influenced by large ur-ban areas like GR01 station in Greece and IT01 close toRome. All models simulate high ozone concentrations overthe Mediterranean Sea; most of them behave satisfactorilyin Malta and Cyprus stations in agreement with the ozoneconcentrations pattern over the seas for the ensemble shownin Fig. 5 and particularly in Malta (Nolle et al., 2002). The



Figure 2. Spatial representations and time variations of the PBL height for the 2009 campaign. Top left panel: mean height of the CHIMEREPBL height issued from ECMWF data. Bottom left panel: mean coefficient of variation for the PBL height. Central panel: hourly variationof the PBL height at the SIRTA station. Top right panel: average diurnal cycle of the PBL height predicted by the models in France. Bottomright panel: average diurnal cycle of the PBL height predicted by the models in Great Britain.

Figure 3. Mean coefficient of variation of the CO concentrationspredicted by the models for the 2006 campaign (no unit). Red colouris assigned for values exceeding the colour scale.

diurnal cycles in Fig. 6 reflect the overall performances de-picted previously. All models fairly simulate the timing ofthe daily peak. For the campaign in 2007, with the exceptionof MINNI, the models overshoot during nighttime and day-time. For the campaign in 2008, the very good shape of theLOTOS-EUROS diurnal cycle is remarkable. For the sum-mertime campaign in 2006, CHIMERE and EMEP provideon average the best diurnal cycles. Focusing on the 2006 and2008 campaigns, the two campaigns which are not biasedby the boundary conditions, LOTOS-EUROS show the bestperformances regarding the bias. For these two campaigns,CAMx has a strong positive bias particularly at night. CAMx

and CHIMERE use exactly the same PBL height of IFS, butnighttime performances of the two models are rather differ-ent. In Fig. 5, the right side is the gridded coefficient of varia-tion that is a standardized measure of the dispersion of modelresults. It is defined as the ratio of the standard deviationto the mean of models. This coefficient is very low for the2006 campaign, below 10 %, and the models have differentresponses along the ship tracks. The coefficients of variationare the highest for the 2007 campaign (Supplement S2) as-sociated with low performances of the ensemble (high nor-malized root mean square errors). France, Spain and Norwayshow the lowest coefficients of variation, indicating a morecoherent behaviour among the models, but not necessarilycorresponding to better model performance than other areas.

At Mace Head (IE31) located on the western part of thedomain, the time series of model results vs. ozone observa-tions shows flat shape for the two winter campaigns with verylow time correlations in 2009 (Fig. 7). The best correlationcoefficients are observed for 2006 and 2008; the models areable to capture the peaks. At this station, the negative biasmentioned in 2009 is roughly the same for LOTOS-EUROS,MINNI and RCG and comparable to the MACC analysis(−20 µg m−3), the other models CAMx, EMEP, CHIMEREand CMAQ have a lower absolute bias (about −10 µg m−3).This behaviour shows that concentrations close to bound-ary conditions are quickly modified, certainly because theregional models restore their own chemical equilibrium inrelation to dynamical processes like deposition and verticaldispersion.



Figure 4. Overall performance of models for ozone, nitrogen dioxide, sulfur dioxide, PM10 and PM2.5 daily mean concentrations for allcampaigns.



Figure 5. Left column: mean ozone concentrations (µg m−3) of the ensemble (ENS) for the 2006 campaign with corresponding observations(coloured dots). Right column: coefficient of variation of models (no unit) constituting the ensemble with corresponding normalized rootmean square errors of the ensemble (coloured dots). Red colour is assigned for values exceeding the colour scale.

Figure 6. Mean ozone diurnal cycles for all campaigns simulated by the models compared with observations. Averaged concentrations areprovided on the right side of the charts.

6.2 Nitrogen dioxide

For NO2, all models perform similarly in terms of correla-tion with values in the range 0.6–0.7 (Fig. 4 and Supple-ment S1). The spatial correlation is much higher in the range0.7–0.9 for all models. Only CMAQ strongly overestimatesthe mean concentrations and CAMx underestimates the con-

centrations for all campaigns. Bessagnet et al. (2014) showedrather low concentrations of elemental carbon compared toother models; this inert species is particularly sensitive tovertical mixing and CAMx presents the highest minimumdiffusion coefficient that is of major importance during stableconditions and partly explains the lower NO2 concentrations.



Figure 7. Time series of hourly concentrations at Mace Head for all models and campaigns

For CAMx, the enhanced mixing influences also O3 concen-trations that are higher than in other models.

The spatial pattern of the ensemble shown for 2009 (Fig. 8)displays high concentrations over the Benelux region, northItaly, the biggest cities and over the shipping tracks. The biasof the ensemble is rather good except for one station in Ser-bia (RS05) with high observed values, probably due to lo-cal sources. The gridded coefficients of variation provided inFig. 8 show that most of differences between models are ob-served over remote areas far from emission regions even iferrors are expected to occur more frequently for low values.As shown for a less reactive species like CO, the differencesof mixing in models over emission areas can lead to largedifferences in modelled concentrations. This effect can beclearly seen over the east Mediterranean for maritime emis-sions where the PBL is different from model to model. Overland, the NO2 chemistry and the different biogenic NO emis-sion modules in the models are believed to explain a large

part of the differences in NO2 concentrations far from urbanareas. As shown in Fig. 8, the root mean square errors of themodels are the highest for the stations close to the emissionareas. The diurnal cycles in Fig. 9 show a general underesti-mation during the afternoon. It should be pointed out that theobserved NO2 concentrations can be slightly overestimated.For some types of analyzers, NO2 is catalytically convertedto NO on a heated molybdenum surface and subsequentlymeasured by chemiluminescence after reacting with ozone.The drawback of this technique is that other oxidized nitro-gen compounds such as peroxyacetyl nitrate and nitric acidare also partly converted to NO (Steinbacher et al., 2007). Inthe observations, the presence of two peaks in NO2 concen-trations is related to the traffic emission peaks occurring inthe morning and the evening. The timing of the peak occur-rences is also modulated by the meteorology: for the 2006and 2008 campaigns performed with identical summer timeshift, we clearly see a time shift of+1 and−1 h, respectively,



Figure 8. Left column: mean nitrogen dioxide concentrations (µg m−3) of the ensemble (ENS) for the 2009 campaign with correspond-ing observations (coloured dots). Right column: coefficient of variation of models (no unit) constituting the ensemble with correspondingnormalized root mean square errors of the ensemble (coloured dots). Red colour is assigned for values exceeding the colour scale.

for the morning and evening peaks corresponding to a laterrise and earlier fall of the PBL. Thus, as expected, the nar-rowest time lag between the two peaks is observed for the2007 campaign. Most of the models predict the first peak tooearly, particularly CHIMERE and CMAQ for the 2006 cam-paign, and the second peak generally occurs too late.

CMAQ shows the strongest nighttime bias that contributesto explain the overall overestimation shown by the model inall campaigns. CMAQ was driven by a different meteorol-ogy that was characterized by very good performance withrespect to both wind speed and PBL height mean bias. Con-versely, IFS-driven models overestimated nighttime windspeed. As nighttime vertical mixing is mainly driven by me-chanical forces, the model results suggest that models tend tounderestimate mixing during stable conditions and, as a con-sequence, IFS-driven models show better results, suggestingcompensation processes.

6.3 Sulfur dioxide

The correlations are rather low for all models in the rangefrom 0.2–0.4 for the 2006 campaign to 0.5–0.6 for the 2007campaign (Fig. 4 and Supplement S1 for all statistics). Twogroups of models are identified CAMx, MINNI and RCGthat largely overestimate the concentrations and CHIMERE,CMAQ, EMEP and LOTOS-EUROS which are closer to theobservations on average with the best performances for theRMSE. The overestimation in the MINNI model could bepartially explained by the low model PBL height. For CAMx,possible reasons such as the vertical distribution of SO2emissions near the harbours and coastal areas, insufficientconversion to sulfate and deposition that was too low werediscussed in Ciarelli et al. (2016). This leads to a positive biasof the ensemble as shown in Fig. 10 (Supplement S4) partic-ularly in western Europe; the normalized RMSE is frequentlyabove 100 % in most parts of Europe. The main hot spots arelocated in eastern Europe, in addition to high concentrations

along the shipping routes. The coefficient of variation is thelowest over emission areas but very high in remote areas likethe oceans far from shipping tracks and over mountain areas.This behaviour, very different from a primary species likeCO, is a first indication of the very different way to simulatethe SO2 chemistry and deposition processes in the models.

The diurnal cycles presented in Fig. 11 show a peak atabout 10:00–12:00 GMT. This peak corresponds with thehourly emission profiles of the industrial sector showing anemission peak at the same hours; however, most of modelspredict a larger decrease in the afternoon. Only CMAQ forthe 2007 campaign captures satisfactorily the diurnal profile.

6.4 PM10

Concerning the RMSE, on average the performances of themodels are similar except CMAQ which has the highest val-ues driven by low correlations and high negative biases par-ticularly for the 2006 campaign (Fig. 4). All models un-derestimate the concentrations generally in the range −3to −10 µg m−3. Except CMAQ, the correlations are in therange 0.4–0.6, but CHIMERE and EMEP reach 0.7 for the2006 campaign. MINNI has the lowest absolute biases forthe 2007, 2008 and 2009 campaigns. The ensemble providesa good picture of the PM10 concentrations in Europe (Fig. 12and Supplement S5) except for two stations (IT01 in Italyand CY02 in Cyprus) with high recorded values. For CY02,high PM10 concentrations are linked to high calcium con-centrations (Bessagnet et al., 2014) due to dust events issuedfrom north Africa. This dust event can be clearly observedfor EMEP in Fig. 14. The spatial patterns show low concen-trations below 5 µg m−3 in remote Scandinavia and three hotspots in the Po Valley, Benelux and south Poland. The coeffi-cient of variation of model results is rather high over the seasand arid areas as well as over areas influenced by biogenicemissions as in Scandinavia. This coefficient is generally thelowest over the western Europe. The best RMSEs of the en-



Figure 9. Mean diurnal cycles of nitrogen dioxide for all campaigns simulated by the models compared with observations. Averaged con-centrations are provided on the right side of the charts.

Figure 10. Left column: mean SO2 concentrations (µg m−3) of the ensemble (ENS) for the 2007 campaign with corresponding observations(coloured dots). Right column: coefficient of variation of models (no unit) constituting the ensemble with corresponding normalized rootmean square errors of the ensemble (coloured dots). Red colour is assigned for values exceeding the colour scale.

semble are observed for the summer campaign in 2006 withvalues below 50 % of the observation data.

EMEP has higher concentrations over north Africa be-cause the model generates dust in this part of the domainand sea salt concentrations are generally higher over the seas.EMEP and CHIMERE perform well for the spatial correla-

tions (Table 5), EMEP captures the high concentrations bet-ter in the south of the domains, whereas CHIMERE performsbetter over the Benelux region (Supplement S5). In 2008,RCG has particularly good spatial correlation compared tothe other models. The missing sea salt emission for CAMx is



Figure 11. Mean SO2 diurnal cycles for all campaigns simulated by the models compared with observations. Averaged concentrations areprovided on the right side of the charts.

Figure 12. Left column: Mean PM10 concentrations (µg m−3) of the ensemble (ENS) for the 2009 campaign with corresponding observations(coloured dots). Right column: coefficient of variation of models (no unit) constituting the ensemble with corresponding normalized rootmean square errors of the ensemble (coloured dots). Red colour is assigned for values exceeding the colour scale.

clearly observed over the ocean with very low PM10 concen-trations impairing the spatial correlations.

As shown in the Supplement S5, most of models underes-timate the highest PM10 concentrations observed in 2008 and2009 by a factor of 2. For the 10 % highest PM10 concentra-tions, MINNI has the lowest underestimations for these two

campaigns, whereas EMEP behaves rather well for the 2006campaign regarding the bias and the correlation. As shownin Bessagnet et al. (2014) the large underestimation in 2009is driven by the underestimation of organic species.

The observed diurnal cycles of PM10 are very flat for allcampaigns with a small peak in the evening (Fig. 13). The



Figure 13. Mean diurnal cycles of PM10 for all campaigns simulated by the models compared with observations. Averaged concentrationsare provided on the right side of the charts.

systematic underestimation of PM10 can be clearly observedbut the shape of the cycle is not very well captured, theevening peak is not reproduced. The models simulate lowconcentrations in the afternoon mainly driven by the eleva-tion of the PBL. For the 2009 campaign, MINNI reproducesthe diurnal cycle very well until 16:00 GMT. As shown inFig. 14, dust concentrations are higher for MINNI in thecentre of the domain. MINNI uses a parameterization forwind-blown dust very productively over any land cover types(Vautard et al., 2005). EMEP mainly produces dust by trafficresuspension and a little over arable land. This higher pro-duction of dust by MINNI in Europe certainly improves thenegative bias for PM usually observed in chemistry transportmodels, particularly in the afternoon when the wind speed ishigher and the soil moisture content lower.

Most of the underestimation of PM10 by the models isdriven by daytime PM10 concentrations that were too low.It is noteworthy that MINNI calculates the lowest PBL thatcould explain its relatively higher PM10 concentrations. Forthe summer campaign in 2006, the PM10 observations showan increase of concentrations in the afternoon while all othermodels tend to predict a decrease, indicating that all modelsare too sensitive to dynamical processes (meteorology) andnot sufficiently sensitive to the chemical formation.

6.5 PM2.5

Performances on PM2.5 concentrations are rather differentcompared to PM10 (Fig. 4). MINNI generally shows a slightpositive bias while all models underestimate the averagedconcentrations, with CMAQ showing the highest negativebias. The performance of CHIMERE on the correlation isvery good for all campaigns, with its RMSE being the low-est for three campaigns. As for PM10, the ensemble cap-tures the spatial patterns of PM2.5 rather well. The concentra-tions in the south of Europe (Fig. 15 and Supplement S6) arenot specifically underestimated except in Cyprus where dustevents also contribute to increase the PM2.5 concentrations.For all campaigns, the coefficient of variation for PM2.5 isthe lowest in Spain but the RMSE of the ensemble is notparticularly low in this region. The coefficient of variationis generally high over the northeast part of the domain. Forall campaigns, the models simulate a hot spot over the northof Italy. As shown in the Supplement S6, CMAQ capturesthe PM2.5 concentrations in Ispra (IT04) for 2007 and 2008campaigns better than the other models. This station, locatedat the border of the Po Valley hot spot, is usually underesti-mated by the models due to the very stably stratified meteoro-logical conditions in this region. The spatial correlations are



Figure 14. Mean dust concentrations (µg m−3) in the PM10 fraction for the 2009 campaign computed by the MINNI, CHIMERE, CAMxand EMEP models.

usually better for PM2.5 for all models except for the summercampaign (Table 5).

As for the PM10 concentrations, the diurnal cycle of PM2.5is rather flat with very small morning and evening peaks(Fig. 16). The models have a different behaviour; they simu-late a sharp decrease of concentrations in the afternoon con-sistent with PM10 diurnal cycles. This confirms the lack ofsecondary production during daytime. The chemical schemesfor the production of organic matter are still incomplete forone main reason. As suggested by Jathar et al. (2014) a largepart of the unspeciated fraction of organic species reactsand produces secondary organic matter and gasoline vehiclescould be an important contributor, as well as wood-burningemissions, according to Denier van der Gon et al. (2015).This unspeciated fraction is not included in our emission in-ventory, explaining a part of the negative bias of models ob-served either in winter or summer campaigns, particularlyduring the afternoon. This suggests that models with nega-tive biases for PM2.5 concentrations are consistent with thelevel of the completeness of our inventory and the state-of-the-art knowledge on SOA modelling.

7 Impact of meteorology on pollutant concentrations

7.1 Impact of the PBL parameterization with MINNIresults for the 2009 campaign

As shown in the previous section, MINNI underestimates thePBL heights calculated at 12:00 GMT from measurementsbut it is in better agreement with hourly data available atSIRTA (Fig. 2). In order to test the effect of PBL heights onair quality predictions, the MINNI model has been run us-ing the PBL from IFS instead of its own parameterization forPBL heights. As shown by Curci et al. (2015), processes inthe PBL can greatly affect the PM2.5 ground concentrations;for instance, temperature and relative humidity can favourthe production of ammonium nitrate in the upper PBL.

Figure 17 shows the average PBL heights and the aver-age concentrations of O3, NO2 and PM10 using MINNI’s pa-rameterizations (left graphs) and the percentage differencebetween the average concentrations calculated with PBLheights given by IFS (PBLIFS) and by MINNI’s parameter-izations (PBLMINNI) (right graphs) using the following for-mula: (PBLIFS−PBLMINNI)/PBLMINNI.

It can be seen that over the seas, on average, PBL heightscalculated with MINNI’s parameterizations (PBLMINNI) arelower than PBL heights given by IFS (PBLIFS) but overland PBLMINNI is higher than PBLIFS in coastal areas, north



Figure 15. Left column: mean PM2.5 concentrations (µg m−3) of the ensemble (ENS) for the 2009 campaign with corresponding observa-tions (coloured dots). Right column: coefficient of variation of models (no unit) constituting the ensemble with corresponding normalizedroot mean square errors of the ensemble (coloured dots). Red colour is assigned for values exceeding the colour scale.

Table 5. PM10 and PM2.5 spatial correlations for all campaigns.

2006 2007 2008 2009

PM10 PM2.5 PM10 PM2.5 PM10 PM2.5 PM10 PM2.5

CAMx 0.58 0.32 0.24 0.60 0.32 0.47 0.07 0.46CHIMERE 0.65 0.32 0.58 0.78 0.39 0.42 0.55 0.66CMAQ 0.50 0.19 0.50 0.80 0.11 0.42 0.11 0.37EMEP 0.75 0.24 0.56 0.62 0.34 0.48 0.68 0.61LOTOS-EUROS 0.34 0.05 0.50 0.61 0.27 0.37 0.50 0.37MINNI 0.61 0.43 0.55 0.58 0.20 0.45 0.32 0.51RCG ND ND ND ND 0.62 0.32 0.44 0.36

Africa, Scandinavian mountains and the middle of the Rus-sian plains, and lower over the rest. Over the sea, PBLIFS arehigher than PBLMINNI more than 50 % while over the landthe differences are between −30 and +30 %.

Figure 17 also shows that the O3 concentrations increase incorrespondence with the increase of PBL heights up to 10 %and more, and decrease where the PBL heights decrease.This behaviour is explained by the fact that with a higherPBL more O3 is entrained from high altitudes where O3 con-centrations are higher than at surface. Since the NO2 sourcesare mainly at surface, the NO2 concentrations generally de-crease with the increase of PBL heights and increase with thedecrease of PBL heights as a consequence of more or less ef-fective dilution, respectively. Over most of Europe, the NO2concentrations decrease up to 8 % when PBLIFS heights areused. The PM10 concentrations respond to PBL height varia-tion in the same way as NO2. The use of PBLIFS heights pro-duces a 4 % decrease of PM10 concentrations in most parts ofEurope but an increase of 6–8 % in coastal areas and Russianplains.

In terms of statistics, the use of the PBL from IFS inMINNI slightly improves the correlations mainly driven byan improvement of time correlations. PM10, PM2.5 and NO2

concentrations are decreased by less than 0.5 µg m−3, im-proving all error statistics reported in Fig. 4f. An increaseof 2.75 µg m−3 is observed for O3 concentrations. It is alsoworth mentioning that the variations in pollutant concentra-tions are small (over the land below 10 % generally) in com-parison to the variations of PBL height; therefore, other fac-tors such as emissions spatial distribution, meteorology (e.g.advection and vertical dispersion, especially in low-wind ar-eas), gas phase chemistry, aerosol physics and chemistryhave to be investigated for improving model performances.

These results clearly show the importance of having goodestimates of PBL heights but they also demonstrate that moreinvestigations are necessary in order to identify the best pa-rameterization of PBL heights as well as vertical diffusivitiesand vertical advection schemes which improve the simulatedconcentrations over all of Europe.

7.2 Influence of meteorology on NO2 concentrationswith CAMx results

Pollutant concentrations are strongly influenced by the re-construction of meteorological fields. In this section, a com-parison of model performances in reproducing wind speedand NO2 concentrations is presented and discussed. Further-


12690 B. Bessagnet et al.: Presentation of the EURODE

Documents

Presentation of the EURODELTA III intercomparison exercise ......B. Bessagnet et al.: Presentation of the EURODELTA III intercomparison exercise 12669 al., 2012), models clearly tend